Many of the claims that 32-bit ARM-based processors could take over the 64-bit server territory of Intel’s high end chips are fanciful. But there is an emerging breed of servers that prioritizes low power and relatively simple, if high performance, tasks – especially those geared at the cloud. They aim to use large numbers of simple processors and this is where ARM licensees like Marvell, STMicro, Nvidia and start-up Calxeda hope to carve out a new market, because they are not up against Intel’s hefty Xeon but against Atom.
Intel remains internally divided on how far low power 32-bit processors can reach up into the server world, but the company is certainly not going to cede even the smallest part of that territory without a fight. It has already helped an Atom-based start-up, SeaMicro, and now, as Calxeda details its 480-core architecture, it is turning its own attentions to ‘microserver’ platforms running on Atoms with power as low as 10W.
Last year Intel took the unusual step of collaborating with a start-up, SeaMicro, to modify one of its own chips, helping the smaller firm to come up with a dual-core Atom design that can deliver 64-bit processing and so run Microsoft software and most server programs. The architecture can support 512 cores. The 64-bit issue is a crucial one ‘ although ARM debuted its A15 core last year, with support for servers, it has been silent on any plans to support 64-bit, which will be a significant brake on progress at the high end.
Dell, which is testing several chipmakers’ wares, said this week that many companies would be reluctant to port software from x86 to ARM or to maintain two software stacks. ‘Fundamentally it’s a software issue,” said the company’s VP of server platforms, Forrest Norrod. “Are there enough benefits from that architecture for porting your code over to that new instruction set … and maintain in two different software stacks? It’s never as trivial as it sounds.”
Other x86 supporters are looking for a place in microservers too, some less friendly to Intel than SeaMicro – VIA’s Nano platform appears in some Dell servers; and AMD plans a version of its low power Bobcat line for the market. And some innovators are even using proprietary designs, notably Tilera with its SoC for massively parallel machines.
In the face of all this, Intel says it will ship its first Atom-based server CPU next year. This will boast power consumption below 10W, one-third of the most efficient current Xeon model, though a 20W Xeon has now gone into production. Intel also plans a 15W server processor based on its new SandyBridge architecture. There is still a way to go to catch ARM-based designs though ‘ Calxeda’s quad-core system-on-chip, running on the Cortex A9, consumes just 5W, which will be attractive to huge cloud operators, which have been using ever more stripped-down boards to try to address their biggest issue, power consumption. Indeed, one of the companies brought out to support Intel at its 10W launch was Facebook ‘ even as Microsoft is said to be one of the most prominent trialists of ARM-based cloud server approaches.
Intel is defining a distinct category from enterprise servers, which it calls microservers, and which are firmly targeted at cloud services. Perhaps with a dose of wishful thinking, given the higher margins on Xeon than Atom, it says this sub-sector will make up less than 10% of the total server space for at least 4-5 years. Microservers consist of a single chassis shared by many single-socket CPU boards and early testers for the Intel design include Dell, Supermicro and Quanta.
The issue for Intel is how to shift its profit model so that it can compete in the microserver space without sacrificing too much margin. Boyd Davis, general manager of Intel’s data center group, said on a conference call to discuss the new product plans: ‘How do we have a category that is optimized enough to address their needs but versatile enough to make it worth our while?’
As Intel looks to diversify its x86 instruction set across a widening range of different server types, Calxeda (formerly SmoothStone) is heading the ARM charge. Last week, it revealed the first solid details of its first ARM-based server chip, which can power platforms with up to 480 cores. The chip runs on a quad-core ARM A9-based processor, so that a low power server could be built with up to 120 CPUs in a compact box. With the inclusion of DRAM, each ARM node will consume an average 5W watts of power during operation and the chip will include a fabric to enable communication between the nodes. Is this the sort of device that future Netflix class services will want to use?
The web server has very specific requirements, and they are tough to address profitably. In the early days, web players needed smaller, cheaper servers than their enterprise counterparts and these were often created by beefing up a desktop processor design. But now they are scaling up and, while some of their tasks may still be simpler than in a corporate data center, in other respects they bring a new level of complexity, especially in communications. The microserver concept includes a fabric inside the box to cut down on ports and handle the networking complexity. Gio Coglitore, director of Facebook Labs, said: ‘We aren’t trying to cram a desktop processor into a server any more ‘ For Facebook there is no one-size-fits-all CPU.’